The local data share, and an operand that was never in any sample

ds_read_b32 and ds_write_b32 translate. The census went from 104 to 105 of 126 and from 4 to 5 of 10 fixtures complete - half the corpus. Sixty-nine execution tests.

The instruction was the easy part; its offset was not. The byte offset lives in the first word, and every earlier probe used the offset-free form - so the field appeared in no sample, and the solved layout had no slot for it. A translator built on that ignores every offset a compiler emits and reads the wrong word, with a perfectly valid address register and nothing to complain about.

Probing with offsets then made the opcode unsolvable, which was the good failure: the reference appends offset:16 to the last operand with no comma, so splitting on commas left v2 offset:16 as one piece that matches no register pattern. The solver now treats a named immediate as the operand it is.

This is the flat accesses' no-base trap again, in a new costume: a probe set that omits a case teaches the solver the case does not exist. Third occurrence. The generalisation already written down after the second one was correct and did not prevent the third, because the shape it takes changes each time - a hidden base, then a hidden width, now a hidden field appended without punctuation.

Surprises.

• The DS opcode sits at bit 17, where the flat families use 18. A test helper with the wrong shift produced a word that decoded as a different local-share instruction, which reported as a missing operand layout rather than as a wrong opcode - so the error named the table rather than the caller.

• Workgroup storage takes one index where a storage buffer takes two. The buffers next door are a struct containing an array and need the member first; this is a bare array and does not. Confusing the two faulted a driver earlier in this crate's life, and writing it correctly here was a matter of remembering that rather than of anything the compiler could check.

Instruction breadth is now exhausted. Every remaining blocker needs either the published instruction set (the three division helpers) or the resource model (exp, MTBUF, MUBUF, MIMG, VINTRP - sixteen of the twenty-one). There is no further unblocked translation work in this thread until a capture arrives.